Custom Resources

InferenceServiceClass

Short names: isc

Scope: Cluster
Version: v1alpha1

• spec

  object

  InferenceServiceClassSpec defines the desired state of InferenceServiceClass.

  The allow-list of API contracts is REQUIRED and non-empty. The rule is stated on the spec rather than on the block that carries the list, because the block itself is optional and a class omitting it would otherwise pass: the refusal has to name the list in both shapes — the block missing, and the block present with an empty list.

  • spec.acceleratorPolicy

    object

    AcceleratorPolicy optionally restricts GPU placement for planner-backed orders.
    • spec.acceleratorPolicy.allowedDeviceClasses

      array of strings

      AllowedDeviceClasses is an optional allow-list of cluster DeviceClass names for planner placement. When empty or omitted, placement considers all DeviceClass candidates in the cluster.
    • spec.acceleratorPolicy.allowedPlacementTypes

      array of strings

      Required value

      AllowedPlacementTypes lists permitted placement modes checked by the planner.
      • spec.acceleratorPolicy.allowedPlacementTypes.Element of the array

        string

        AcceleratorPlacementType selects whole-device or partition placement.

        Allowed values: Partition, WholeDevice

    • spec.acceleratorPolicy.allowedSharingModes

      array of strings

      Required value

      AllowedSharingModes lists permitted sharing modes checked by the planner.
      • spec.acceleratorPolicy.allowedSharingModes.Element of the array

        string

        AcceleratorSharingMode selects how accelerators may be shared between workloads.

        Allowed values: Shared, Dedicated

    • spec.acceleratorPolicy.maxAcceleratorCount

      integer

      Required value

      MaxAcceleratorCount is the upper bound on the number of devices a launch plan may assign to one replica. The order names no count of its own, so this bounds an outcome of planning rather than a request.

      Allowed values: 1 <= X

    • spec.acceleratorPolicy.maxSharePercent

      integer

      MaxSharePercent is the optional upper bound for accelerator share per requested device.

      Allowed values: 1 <= X <= 100

    • spec.acceleratorPolicy.minSharePercent

      integer

      MinSharePercent is the optional lower bound for accelerator share per requested device.

      Allowed values: 1 <= X <= 100

  • spec.admissionPolicy

    object

    AdmissionPolicy optionally restricts namespaces that may use this class.
    • spec.admissionPolicy.allowedNamespaces

      array of strings

      Required value

      AllowedNamespaces lists namespaces permitted to use this class. The value “*” allows any namespace.
  • spec.exposurePolicy

    object

    Required value

    ExposurePolicy defines endpoint exposure and authentication defaults.
    • spec.exposurePolicy.authentication

      string

      Required value

      Authentication selects how clients authenticate to the inference API.

      Allowed values: Token, None

    • spec.exposurePolicy.https

      object

      HTTPS optionally configures Ingress TLS when type is External.
      • spec.exposurePolicy.https.certManager

        object

        CertManager names the ClusterIssuer when mode is CertManager.
        • spec.exposurePolicy.https.certManager.clusterIssuerName

          string

          Required value

          ClusterIssuerName is the cert-manager ClusterIssuer name.

          Minimal length: 1

      • spec.exposurePolicy.https.customCertificate

        object

        CustomCertificate names a TLS secret in d8-system when mode is CustomCertificate.
        • spec.exposurePolicy.https.customCertificate.secretName

          string

          Required value

          SecretName is the TLS secret name in the d8-system namespace.

          Minimal length: 1

      • spec.exposurePolicy.https.mode

        string

        Mode selects the HTTPS usage mode for External exposure.

        Allowed values: Disabled, CertManager, CustomCertificate, OnlyInURI

    • spec.exposurePolicy.type

      string

      Required value

      Type selects cluster-local or external exposure.

      Allowed values: ClusterLocal, External

  • spec.modelPolicy

    object

    ModelPolicy restricts catalog model orders. Its allow-list of API contracts is required and non-empty; its restrictions on the format and the size of a model stay optional.
    • spec.modelPolicy.allowedEndpointTypes

      array of strings

      Required value

      AllowedEndpointTypes is the allow-list of inference API contracts for orders using this class. Required and non-empty: the contract of an order is picked from this list, and an order names none.
      • spec.modelPolicy.allowedEndpointTypes.Element of the array

        string

        InferenceAPIEndpointType is the serving API taxonomy aligned with ai-models catalog facts.

        Allowed values: Chat, Embeddings, Rerank

    • spec.modelPolicy.allowedFormats

      array of strings

      AllowedFormats optionally restricts artifact formats from catalog facts.
    • spec.modelPolicy.maxParameterCount

      string

      MaxParameterCount optionally caps model size (for example 32B, 70B).

      Pattern: ^[0-9]+(\.[0-9]+)?[MmBb]$

  • spec.modelStorageClassName

    string

    ModelStorageClassName optionally selects the StorageClass of the model volume of the orders of this class. Nobody computes it: it is a decision about WHERE the model lies, and it belongs to the administrator of the class rather than to the owner of an order. Omitted leaves the volume to the default StorageClass of the cluster, and the platform adds no second default of its own.
  • spec.scalingPolicy

    object

    ScalingPolicy defines the replica bounds of the orders of this class and the PriorityClass names it offers them. It is the only source of those bounds — an order names none of its own — so a class without this block gives its orders one replica.
    • spec.scalingPolicy.allowedPriorityClassNames

      array of strings

      AllowedPriorityClassNames lists the Kubernetes PriorityClass names this class offers its orders. An order names none of its own, so a non-empty list decides: its first entry is the effective name. Empty or omitted leaves the effective name to the cluster globalDefault PriorityClass.
    • spec.scalingPolicy.maxReplicas

      integer

      MaxReplicas is the maximum replica count for orders of this class.

      Allowed values: 1 <= X

    • spec.scalingPolicy.minReplicas

      integer

      MinReplicas is the minimum replica count for orders of this class. The schema minimum is 1 on this user-facing input; capacity preemption writes bounds on the order workload and may drive it to 0.

      Allowed values: 1 <= X

  • spec.updatePolicy

    object

    UpdatePolicy optionally defines how the workload of an order is updated.
    • spec.updatePolicy.strategy

      string

      Required value

      Strategy names the way the replicas of an order are moved when the platform may move them.

      Allowed values: RollingUpdate

• status

  object

  InferenceServiceClassStatus defines the observed state of InferenceServiceClass.
  • status.conditions

    array of objects

    Conditions report validation and readiness.
    • status.conditions.lastTransitionTime

      string

      lastTransitionTime is the last time the condition transitioned from one status to another. This should be when the underlying condition changed. If that is not known, then using the time when the API field changed is acceptable.
    • status.conditions.message

      string

      message is a human readable message indicating details about the transition. This may be an empty string.

      Maximum length: 32768

    • status.conditions.observedGeneration

      integer

      observedGeneration represents the .metadata.generation that the condition was set based upon. For instance, if .metadata.generation is currently 12, but the .status.conditions[x].observedGeneration is 9, the condition is out of date with respect to the current state of the instance.

      Allowed values: 0 <= X

    • status.conditions.reason

      string

      reason contains a programmatic identifier indicating the reason for the condition’s last transition. Producers of specific condition types may define expected values and meanings for this field, and whether the values are considered a guaranteed API. The value should be a CamelCase string. This field may not be empty.

      Length: 1..1024

      Pattern: ^[A-Za-z]([A-Za-z0-9_,:]*[A-Za-z0-9_])?$

    • status.conditions.status

      string

      status of the condition, one of True, False, Unknown.

      Allowed values: True, False, Unknown

    • status.conditions.type

      string

      type of condition in CamelCase or in foo.example.com/CamelCase.

      Maximum length: 316

      Pattern: ^([a-z0-9]([-a-z0-9]*[a-z0-9])?(\.[a-z0-9]([-a-z0-9]*[a-z0-9])?)*/)?(([A-Za-z0-9][-A-Za-z0-9_.]*)?[A-Za-z0-9])$

  • status.observedGeneration

    integer

    ObservedGeneration is the most recent metadata generation observed by the controller.
  • status.phase

    string

    Phase is Pending, Ready, or Failed.

InferenceService

Short names: is

Scope: Namespaced
Version: v1alpha1

• spec

  object

  InferenceServiceSpec defines the desired state of InferenceService.
  • spec.inferenceServiceClassName

    string

    Required value

    InferenceServiceClassName selects the cluster InferenceServiceClass.

    Minimal length: 1

  • spec.model

    object

    Required value

    Model describes the model source and reference. Together with the class reference it is the whole of an order: the inference API contract is picked from the allow-list of the class, and the runtime and its parameters are named by the recipe of the launch plan.
    • spec.model.authSecretRef

      object

      AuthSecretRef references a Secret with Hugging Face credentials, on the DIRECT path only: the artifact of a catalog entry is mounted by the catalog module, which holds its own credentials, so on that path the field is refused on write.
      • spec.model.authSecretRef.key

        string

        Key is the Secret data key holding the credential. When omitted, the platform uses “token”.
      • spec.model.authSecretRef.name

        string

        Required value

        Name is the Secret metadata.name in the same namespace.

        Minimal length: 1

    • spec.model.parameterCount

      string

      ParameterCount declares model size on the DIRECT path, and only there: an order through the catalog takes the size from the facts of the catalog entry, and the field is refused on write.

      Pattern: ^[0-9]+(\.[0-9]+)?[MmBb]$

    • spec.model.ref

      object

      Required value

      Ref is a structural object: HuggingFace uses name only; ai-models uses kind+name (+ optional namespace). Discriminated by src via CEL on this struct.
      • spec.model.ref.kind

        string

        Kind is Model or ClusterModel for catalog orders; omitted for HuggingFace.

        Allowed values: Model, ClusterModel

      • spec.model.ref.name

        string

        Required value

        Name is the Hugging Face model id or catalog CR metadata.name.

        Minimal length: 1

      • spec.model.ref.namespace

        string

        Namespace is optional for kind Model and must match the InferenceService namespace.
    • spec.model.src

      string

      Required value

      Src selects HuggingFace or ai-models.

      Allowed values: HuggingFace, ai-models

• status

  object

  InferenceServiceStatus defines the observed state of InferenceService.
  • status.authSecretName

    string

    AuthSecretName is the generated token Secret when the class uses Token authentication.
  • status.conditions

    array of objects

    Conditions report class resolution, model resolution, order workload readiness and service readiness.
    • status.conditions.lastTransitionTime

      string

      lastTransitionTime is the last time the condition transitioned from one status to another. This should be when the underlying condition changed. If that is not known, then using the time when the API field changed is acceptable.
    • status.conditions.message

      string

      message is a human readable message indicating details about the transition. This may be an empty string.

      Maximum length: 32768

    • status.conditions.observedGeneration

      integer

      observedGeneration represents the .metadata.generation that the condition was set based upon. For instance, if .metadata.generation is currently 12, but the .status.conditions[x].observedGeneration is 9, the condition is out of date with respect to the current state of the instance.

      Allowed values: 0 <= X

    • status.conditions.reason

      string

      reason contains a programmatic identifier indicating the reason for the condition’s last transition. Producers of specific condition types may define expected values and meanings for this field, and whether the values are considered a guaranteed API. The value should be a CamelCase string. This field may not be empty.

      Length: 1..1024

      Pattern: ^[A-Za-z]([A-Za-z0-9_,:]*[A-Za-z0-9_])?$

    • status.conditions.status

      string

      status of the condition, one of True, False, Unknown.

      Allowed values: True, False, Unknown

    • status.conditions.type

      string

      type of condition in CamelCase or in foo.example.com/CamelCase.

      Maximum length: 316

      Pattern: ^([a-z0-9]([-a-z0-9]*[a-z0-9])?(\.[a-z0-9]([-a-z0-9]*[a-z0-9])?)*/)?(([A-Za-z0-9][-A-Za-z0-9_.]*)?[A-Za-z0-9])$

  • status.constraints

    object

    Constraints publishes effective replica bounds and catalog scope after validation.
    • status.constraints.maxReplicas

      integer

      MaxReplicas is the effective maximum replica count, and it has two origins: outside a preemption episode it is the bound the validation of the order against the scaling policy of its class produced, and while an episode holds the order as a donor it is the bound of that hold, which the restore of the donor puts back to the policy bound.
    • status.constraints.minReplicas

      integer

      MinReplicas is the effective minimum replica count.
    • status.constraints.modelScope

      string

      ModelScope is the catalog model visibility from catalog facts.
    • status.constraints.priorityClassName

      string

      PriorityClassName is the effective Kubernetes PriorityClass name after validation. Empty when the order uses the cluster default that is absent (numeric priority 0).
  • status.endpoint

    object

    Endpoint publishes the resolved inference API URL.
    • status.endpoint.type

      string

      Type mirrors exposure semantics (ClusterLocal or External).
    • status.endpoint.url

      string

      URL is the HTTP(S) endpoint clients use to call the inference API.
  • status.model

    object

    Model publishes catalog path metadata after successful validation of the order against its class.
    • status.model.endpointType

      string

      EndpointType publishes the inference API type the platform selected for the order after validating it against its class. An order names no type: the platform takes the single type the class allows, or the one type the allow-list of the class and the facts of the model have in common.

      Allowed values: Chat, Embeddings, Rerank

  • status.observedGeneration

    integer

    ObservedGeneration is the most recent metadata generation observed by the controller.
  • status.phase

    string

    Phase is Pending, Ready, Degraded, or Failed.
  • status.resolved

    object

    Resolved publishes the placement the order runs on after successful planner reconciliation.
    • status.resolved.acceleratorMemoryGiB

      integer

      AcceleratorMemoryGiB is usable accelerator memory for the selected placement.
    • status.resolved.acceleratorProductName

      string

      AcceleratorProductName is the product name of the selected accelerator.
    • status.resolved.deviceClass

      string

      DeviceClass is the DRA device class the planner recommends for the device claim of the order.
    • status.resolved.excludedDeviceClasses

      array of strings

      ExcludedDeviceClasses lists cluster DeviceClass CR names excluded after failed replans.
    • status.resolved.placementMode

      string

      PlacementMode is the selected placement mode (Partition or WholeDevice).

      Allowed values: Partition, WholeDevice

    • status.resolved.replanCount

      integer

      ReplanCount is the number of ResourceClaim-driven replans performed.
    • status.resolved.sharePercent

      integer

      SharePercent is the share of the device the plan ASSIGNED to this order. A hundred means the whole device. It belongs to the published placement on both grounds of the closed composition: the projection of occupancy of the platform reads it back — without it a card holding a neighbour at thirteen percent looks either free or fully taken — and the owner of the order sees by it how much of the card is theirs.
    • status.resolved.sharingMode

      string

      SharingMode is the selected sharing mode when applicable, independent of the placement mode. A partition of an accelerator may be held alone or shared with neighbours.

      Allowed values: Shared, Dedicated